GB300 NVL72 multi rack AI factory

A GB300 NVL72 AI factory joins two or more NVL72 racks into one system. Each rack is a 72 GPU NVLink domain with 20 TB of HBM3e and 130 TB/s of NVLink bandwidth as published by NVIDIA, and the racks connect over 800 Gb/s ConnectX-8 ports into a rail optimized Quantum-X800 or Spectrum-X fabric. We design the scale out fabric, storage and control plane with the order so the factory is deployable as delivered.

It is the platform for reasoning model inference at scale, where long context and test time scaling reward a large NVLink domain, and for training runs that would otherwise be bound by inter node traffic. The supporting tiers follow our standard rules: one terabyte of usable flash per GPU under a parallel file system with GPUDirect Storage, and six control plane nodes for scheduling, fabric management, monitoring and provisioning.

Every GB300 NVL72 project starts with a site survey, because the racks are fully liquid cooled and need a facility water loop, coolant distribution units and a high density power feed. NVIDIA has not published a rack level power figure, so power is planned with the site survey and the system partner. Racks are delivered in phases to your own hall or to a colocation facility with the required water loop and power density, and each is commissioned and tested before the next is energized.